River planform change and lateral channel migration
Multi-epoch optical imagery, anchored by the Landsat archive, lets analysts map lateral channel migration, cutoff events, and braiding-pattern shifts across decades. The same record exposes unreported human interventions that alter natural migration rates.
Sensors
- Landsat 1–9 (USGS/NASA): 30 m multispectral resolution (bands 1–7 on OLI), 16-day revisit on a single satellite, archive continuous from 1972. The only freely available record long enough to document decadal migration rates and rare avulsion events.
- Sentinel-2 MSI (ESA Copernicus): 10 m resolution in visible and near-infrared bands, 5-day revisit at mid-latitudes with both satellites. Near-infrared band 8 (842 nm) gives sharp water-sediment contrast for bankline delineation on rivers wider than roughly 20–30 m.
- Planet SuperDove: 3–5 m resolution, near-daily revisit over most land areas. Resolves channel threads in braided systems that Sentinel-2 blurs together, though the commercial archive is shallow compared with Landsat and cloud cover remains an unresolved constraint.
- SWOT KaRIn (NASA/CNES, launched 2022): Ka-band radar interferometer measuring water-surface width and elevation at roughly 50–100 m cross-track resolution, 21-day repeat. Adds a hydraulic dimension to planform mapping: width changes detected optically can be cross-checked against water-surface slope from the same overpass.
What a floating roof gives away
Water absorbs near-infrared radiation strongly. Dry sand and gravel reflect it. That physical contrast, not any clever algorithm, is the foundation of satellite-based bankline mapping. In a Landsat or Sentinel-2 near-infrared band, the boundary between an active channel and a sediment bar is often the sharpest edge in the scene. Automated thresholding on the Normalised Difference Water Index (NDWI, using green and near-infrared bands) can delineate banklines to within one or two pixels of their true position under clear-sky conditions.
The practical floor matters. At 30 m Landsat resolution, channels narrower than roughly 30–60 m are systematically underestimated in width because mixed pixels straddle the edge. Sentinel-2 at 10 m pushes that floor to around 15–20 m. Planet SuperDove at 3–5 m resolves individual braid threads in gravel-bed rivers that the coarser sensors treat as a single channel. No optical sensor sees through cloud, and tropical rivers, which tend to migrate fastest, sit under persistent cloud cover for months at a time. Radar-based flood mapping (covered on a sibling page) can fill some of those gaps, but radar backscatter from turbid braided channels is ambiguous in ways that near-infrared contrast is not.
Fifty years of lateral drift, read from a single archive
The Landsat programme launched its first satellite in 1972. That is not a marketing claim; it is a physical fact with direct consequences for hydrology. A river that migrates at 10 metres per year will have moved 500 metres since the first Landsat overpass. At 30 m resolution, that displacement is unambiguous across 16 or more image epochs. Published studies of the Brahmaputra, the Ucayali, and the Lower Mississippi have used this archive to document migration rates ranging from a few metres per year in bedrock-confined reaches to more than 200 m per year in unconstrained alluvial plains.
The analysis is not simply a matter of overlaying images. Cloud masking, seasonal water-level variation (a river looks wider in flood than in low flow), and sensor cross-calibration across Landsat generations all introduce noise. The standard approach is to build a time series of cloud-free composites, apply consistent NDWI thresholds, extract centreline or bankline vectors per epoch, and then compute migration vectors between successive epochs. Cutoff formation, where a meander loop is abandoned and an oxbow lake forms, appears as a discrete topological change in the centreline graph and is straightforward to detect automatically. Avulsions, where a river abruptly shifts to a new course, produce migration vectors far larger than the surrounding background rate and stand out clearly.
When the river stops moving and nobody announced it
Natural migration rates are not constant. They respond to flood frequency, sediment supply, and vegetation on the floodplain. They also respond to human intervention: bank revetment with riprap or concrete, gravel extraction from the active channel, and upstream dams that trap sediment and starve downstream reaches of the material that drives lateral erosion. These interventions are frequently unreported, particularly in jurisdictions with limited environmental governance.
Satellite time series expose the signature. A reach that migrated at 15 m per year through the 1990s and then abruptly stopped in 2003 has almost certainly been engineered. Gravel extraction leaves a different mark: the channel widens and the thalweg deepens, which shows up as increased water-surface area at a given discharge. Neither of these changes requires fieldwork to detect. The archive is the witness.
This matters for sediment-budget studies because a revetted reach exports sediment to the reach downstream without replenishment from lateral erosion, altering the downstream morphology in ways that propagate over decades. It also matters for infrastructure planning: a bridge designed for a stable channel may face unexpected scour if upstream extraction accelerates bed incision.
Flood risk is partly a geometry problem
A river that has migrated 300 m toward a road embankment over 30 years is not an unpredictable hazard. It is a documented trend. Migration-rate maps derived from the Landsat archive can be extrapolated forward, with appropriate uncertainty bounds, to estimate when a channel centreline will reach a given asset. The uncertainty grows with the forecast horizon and with any change in the upstream sediment or flow regime, so these projections are best treated as conditional: given the migration rate observed over the last two decades, the channel will reach the embankment within X to Y years unless something changes.
SWOT adds a useful cross-check. Because KaRIn measures water-surface width directly from radar, it is independent of the optical cloud problem and independent of the analyst's NDWI threshold choice. Where SWOT width observations agree with optically derived banklines, confidence in the planform mapping is higher. Where they diverge, the discrepancy usually points to a calibration issue or to a reach where dense riparian vegetation is obscuring the true bankline in the optical image.
What the method cannot do
Optical bankline mapping is a surface measurement. It says nothing about bank height, bank material, or the depth of scour that determines whether a bank will fail. A channel that has not migrated laterally can still be incising vertically, which is a different and sometimes more dangerous process. Vertical change requires either field survey, lidar, or the kind of water-surface elevation data that SWOT provides, and even SWOT's 50–100 m cross-track resolution is coarse for narrow channels.
Cloud cover is the chronic limit. In the Amazon basin or the Irrawaddy delta, finding even one cloud-free Landsat scene per year for a specific reach is not guaranteed. Annual compositing helps but blurs the timing of rapid events. Planet's near-daily revisit improves the situation for commercial clients, but the archive only extends back to 2016, which is insufficient for decadal trend analysis on its own.
Satellize runs multi-epoch bankline analysis on open Landsat and Sentinel-2 archives, with optional Planet tasking for recent high-resolution epochs, and delivers migration-rate maps as GIS layers with per-reach uncertainty estimates. The Tonga crop-estimation programme demonstrated the same compositing and change-detection pipeline in a different domain; the underlying approach transfers directly to fluvial planform work.
Typical figures
| Spatial resolution (bankline mapping) | 30 m (Landsat), 10 m (Sentinel-2), 3–5 m (Planet SuperDove) |
| Minimum detectable channel width | ~30–60 m at Landsat resolution; ~15–20 m at Sentinel-2; ~6–10 m at Planet SuperDove |
| Revisit period | 16 days per Landsat satellite; 5 days (Sentinel-2A+B combined); near-daily (Planet) |
| Archive depth | Landsat: 1972 to present (longest free global archive); Sentinel-2: 2015 to present; Planet: 2016 to present |
| Key spectral bands | Near-infrared (Landsat Band 5 ~865 nm; Sentinel-2 Band 8 ~842 nm) for water-sediment contrast; green band for NDWI computation |
| SWOT KaRIn resolution and repeat | ~50–100 m cross-track, 21-day repeat; measures water-surface width and elevation |
| Cloud cover limitation | Optical sensors blind under cloud; annual clear-sky composites may be unavailable in humid tropics |
| Typical migration-rate detection floor | ~1–2 pixels per epoch; roughly 30–60 m per interval at Landsat, 10–20 m at Sentinel-2 |
| Deliverable formats | GeoPackage or Shapefile bankline vectors, GeoTIFF migration-rate rasters, per-reach summary CSV |
Analytics Satellize can run
| Multi-epoch bankline vector series | NDWI thresholding on cloud-masked composites, per-epoch centreline extraction | GeoPackage of dated bankline polygons, one layer per epoch, spanning available archive |
| Lateral migration rate map | Centreline displacement vectors between successive epochs, smoothed with robust regression to reduce cloud-noise artefacts | GeoTIFF raster and per-reach CSV of mean migration rate (m/yr) with 90% confidence interval |
| Cutoff and oxbow detection log | Topological change detection in centreline graph across epochs, flagging loop-closure events | Point GIS layer of cutoff events with date range, abandoned channel area, and oxbow persistence score |
| Human-intervention anomaly report | Comparison of pre- and post-intervention migration rates; step-change detection in time series | PDF report identifying reaches where migration rate changed abruptly, with candidate cause classification (revetment, extraction, dam) |
| Infrastructure proximity forecast | Linear and trend-extrapolated migration projections from observed rate, with uncertainty bounds | GIS layer showing projected bankline positions at 5, 10, and 20-year horizons with confidence envelopes |
| Channel width time series (SWOT cross-check) | SWOT KaRIn width observations co-registered to optical bankline epochs for consistency validation | Per-reach width time series combining optical and radar-derived estimates, flagging discrepancies |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.